Book a meeting
All articles
L&D & Capability19 November 2026·6 min read

Only 6% of Training Studies Reported a Needs Analysis. A Preference Survey Is Not One.

In the largest meta-analysis of workplace training, only 22 of 397 effect sizes came from studies that documented a needs analysis. The course-preference survey that has replaced it answers none of the three questions a real needs analysis asks.

By AssessAll Editorial

A training needs analysis is the diagnostic step that establishes whether a performance gap exists, whether training can close it, and who specifically needs it. It has three levels — organisational, task and person. A survey asking employees which courses they want is none of these. It measures preference, not capability.

The evidence base has a hole in exactly this shape

The largest meta-analysis of workplace training effectiveness pooled 397 independent effect sizes drawn from 162 sources, covering 33,325 trainees. Training worked, and worked consistently: the sample-weighted mean effect was d = 0.62 overall, holding up at the reaction (d = 0.60), learning (d = 0.63), behavioural (d = 0.62) and results (d = 0.62) levels.

The notable figure is elsewhere. Only 6% of those data points — 22 of 397 — came from studies that reported having conducted a needs assessment at all (Arthur, Bennett, Edens & Bell, 2003).

With 22 data points, the authors could not establish much. They compared outcomes by which levels of needs assessment had been performed and found "no clear pattern of results." That is the honest state of the evidence. We cannot demonstrate from this meta-analysis that needs analysis improves training outcomes, because almost nobody writes down that they did one. The hole may be in the reporting rather than the practice — but a step nobody documents is a step nobody can audit, and in most L&D functions that documentation is a spreadsheet of course requests.

Three levels, and a preference survey answers none of them

Salas, Tannenbaum, Kraiger and Smith-Jentsch, reviewing the field for Psychological Science in the Public Interest, call needs analysis "the first and probably the most important step" in training design, and observe that systematic analysis "is often skipped or replaced by rudimentary questions." Their framework has three components, and each answers a question the others cannot.

Organisational analysis

Does this capability matter to the strategy, and is the workplace ready to absorb it? This covers culture, norms, resources and managerial support. The authors note that designers and researchers alike "often overlook this strategic component."

Task analysis

What are the critical work functions, and which knowledge, skills and abilities do they actually require? Salas and colleagues call this "the blueprint for the training," and recommend separating what people need to know from what they need to be able to access — a distinction that quietly removes a large share of most content libraries.

Person analysis

Who specifically lacks the capability the task analysis identified? With finite budget, the recommendation is to target "those with the largest gaps between actual and needed competencies" — which requires knowing the size of each gap, per person.

A course-preference survey asks a fourth question entirely: what would you like to learn? The answer tracks interest, perceived career value and whatever a manager mentioned last quarter. Citing Baddeley and Longman's 1978 training research, Salas and colleagues note that employees are frequently unable to say what training they genuinely need. That is not a failing of the people answering. It is a property of the instrument.

Roughly half of what decides transfer is not inside the trainee

A rigorous person analysis is still only part of the diagnosis. A meta-analysis of training transfer covering 89 sources, 325 independent correlations and 12,496 participants found person-side and situation-side predictors in the same range (Blume, Ford, Baldwin & Huang, 2010).

Person-side predictors (corrected correlations with transfer)

  • Cognitive ability: ρ = .37 (k = 10, N = 1,452), though almost entirely from laboratory studies
  • Voluntary participation: ρ = .34 (k = 5, N = 1,413)
  • Conscientiousness: ρ = .28 (k = 5, N = 433)
  • Motivation: ρ = .23 (k = 24, N = 2,574)

Situation-side predictors (corrected correlations with transfer)

  • Transfer climate: ρ = .27 (k = 8, N = 870)
  • Work environment overall: ρ = .23 (k = 22, N = 2,085)
  • Supervisor and peer support: ρ = .21 (k = 12, N = 1,075)

The authors' summary is blunt: "there is no clear superiority of individual variables over situational variables, or the reverse." A needs analysis that examines only people is therefore blind to about half of what determines whether the training changes anything — and the organisational-analysis level is precisely the step that would have caught it.

Note the voluntary-participation figure too. Choice predicts transfer at ρ = .34, which makes the common sequence — survey people on what they want, then mandate a different curriculum — a poor use of the one lever the survey did hand you.

Sometimes the correct finding is "do not train"

The most useful output of a needs analysis is occasionally a decision not to run a programme. Salas and colleagues state plainly that training "is not always the ideal solution to address performance deficiencies." Their review also reports, citing Tannenbaum's 1997 work, that only 7–9% of skill acquisition in organisations comes from formal training at all. At the time of that review, US organisations spent roughly US$135 billion a year on training.

If the constraint is a broken process, a missing tool, an incentive pointing the other way or a manager who penalises the new behaviour, the course will measure as a failed intervention while the real cause goes unexamined. A needs analysis incapable of returning "do not train" is not a diagnosis. It is a procurement form.

How to run a defensible one in a week

  1. Start from a business metric, not the catalogue. Name the number that should move — first-call resolution, error rate, cycle time, conversion — before naming any capability. If no metric moves, the organisational analysis has already answered the question.
  2. Write the task down before the content. Three or four subject-matter experts, two hours, one list of critical functions and the knowledge and skills each requires. Mark each item need-to-know or need-to-access.
  3. Measure the gap rather than surveying it. Person analysis requires a score per person on the capability the task analysis named, not a self-reported confidence rating. A short scenario-based or role-relevant assessment gives you gap sizes you can rank and a baseline you can re-test against later; on AssessAll, AI-graded scenario assessments run on pay-as-you-go credits at ₹30 (US$0.50) each, with 100 free individual credits or 250 corporate credits on signup, so a one-team diagnostic needs no licence negotiation.
  4. Audit the environment before you buy the programme. Ask the trainees' managers two questions: will this behaviour be visible in the work, and what currently happens to someone who tries it? Transfer climate predicts transfer about as strongly as conscientiousness does.
  5. Fix the criterion and the date in advance. Decide now which measure will be read, at what interval, and what result would count as success. Deciding afterwards is how a programme becomes unfalsifiable.
  6. Re-measure the same construct, not a satisfaction form. The post-training measure has to be comparable to the baseline, which means the same construct and a defensible scoring rule.

One caution: if a capability assessment also gates a promotion, a posting or a bonus, it has become an employment-decision tool, and the evidentiary standard rises accordingly — see the US EEOC's guidance on employment tests and selection procedures and the *Standards for Educational and Psychological Testing*. A development framing does not exempt a decision-making instrument.

When a preference survey is the right instrument

The survey is the right instrument for several real jobs.

  • Allocating an elective catalogue. If attendance is voluntary and the budget is discretionary, demand data is exactly what you need.
  • Preserving the choice effect. Voluntary participation predicted transfer at ρ = .34. Where you can let people choose, the choosing itself is doing work.
  • Surfacing what the task analysis missed. Open-text responses are a cheap way to find emerging tasks that your experts did not mention.

The error is not running the survey. The error is filing it under needs analysis and skipping the three levels that would have told you whether to train, what to train and whom.

The takeaway

In the largest meta-analysis of workplace training, 94% of effect sizes came from studies that never documented a needs analysis — and the practice that has replaced it answers a question about preference rather than capability. Run the organisational, task and person levels properly, measure the gap instead of asking about it, and accept that a well-run diagnosis will sometimes tell you not to train at all.

#training-needs-analysis#l-and-d-measurement#capability-measurement#training-transfer#person-analysis

Measure it, don't guess it.

Start free with 100 credits — or write to solutions@bodhih.com.

Start free